Been using Omma (by the Spline team, omma.build) and it’s genuinely insane. One prompt gives you scroll-driven animations, custom shaders, 3D models, all generated in parallel by multiple AI agents simultaneously. Check the community page, the outputs are wild.
Which got me thinking, for pure portfolio / visual websites, do we even need a Three.js dev anymore?
And if AI is taking over that space, are domains like virtual tours, 3D education platforms, medical visualizations, architectural viewers literally the only areas left where real Three.js engineering still matters? Or am I missing something, are there other spaces where 3D on the web is still deeply underexplored?
Threejs isn’t a “pure portfolio/visual website” tool.. it’s an api. A scenegraph built on webgl/gpu.
If an AI can come up with a better abstraction over WebGL then threejs, that would be cool too, but as it stands.. if you ask an LLM to make a cool 3d effect on the web.. dollars to donuts it will import threejs (v160 for some reason related to the knowlege cutoff) to do it.
And what do you do if you need to engineer something? Trust me, you can’t prompt your way out of it without solid skills, coding and engineering skills included.
Just try prompting a full commercial app and see how it goes. I tried it, got about 95% done and I was really impressed, then got stuck. The code was irreparable at that point, so I had to quit after spending about $200 and an entire week prompting, fixing one thing only to break another, tyring to fix an issue that I had it created so many 'fixes" that the code was a nightmare literally, ina single function it create about twnty boolean flags… at that point it was too late to fix anything in there.
By now, every good developer should know that prompting should be done sequentially. You need to understand what’s happening in order to properly architect the app, bring it to completion, and not lose your mind in the process.
Vibe coding can be impressive, but as soon as you try to build a full project with all the small details involved, it fails badly. I mean, it fails so badly that it makes you never want to use it again.
So yes at this point agentic coding is the way but you still need to be a good developer and engeener to make any sense of it!
The AI is right to pick three best choice
the reason for v160 is how this version is served across the web. You can push it at 170+, but it start to use strange formating (like for TSL nodes), and it may have to read the new documentation (cost more tokens).
AI is also very volatile on webGPU. Mostly because it doesn’t know how to debug it. Sometimes I had Claude rewriting my code to GLSL, verify, then convert it back.
So on a pure technical perspective, LLM are not in the future. They’re 20 versions late for 3D on the web. It’s possible to vibe code on the bleeding edge with WGSL / TSL and heavy instancing/optimisation , but this take extra care/time and costs due to instability.
Yeah, totally agree. Once things move beyond surface-level visuals, you hit a wall pretty quickly without solid engineering fundamentals and domain knowledge.
It’s becoming very easy to generate portfolio-style websites now, but the moment you deal with real-world complexity state, performance, edge cases, integrations that’s where actual Three.js expertise still matters.
Feels less like “AI replacing devs” and more like it compresses the easy layer, while the harder, niche problems still need real depth.
Yeah, especially with WebGPU it’s still pretty rough for LLMs.
At that level, you really need actual expertise to reason about performance, debugging, and architecture. AI can assist, but it’s definitely not leading there yet.
best choice